QuerySet.values_list() combined with .extra() or .annotate() may produce wrong .union()

[master]

Regression introduced in 2.0, still present in 2.0.3.

Easy context to reproduce the problem: suppose a Message model, in a postman app. Only one ordinary field is enough for the demo.

qs1 = Message.objects.extra(select={'count': 0}).values_list('id', 'count').order_by()
print(qs1.query)  # as expected:
# SELECT (0) AS "count", "postman_message"."id"
# FROM "postman_message"

qs2 = Message.objects.values('somefield').annotate(count=models.Count('pk')).annotate(id=models.Max('pk')).values_list('id', 'count').order_by()
print(qs2.query)  # as expected:
# SELECT COUNT("postman_message"."id") AS "count",  MAX("postman_message"."id") AS "id"
# FROM "postman_message" GROUP BY "postman_message"."somefield"

print(qs1.union(qs2).query)  # !! WRONG !! the qs2 part is truncated:
# SELECT (0) AS "count", "postman_message"."id" FROM "postman_message"
# UNION
# SELECT MAX("postman_message"."id") AS "id" FROM "postman_message" GROUP BY "postman_message"."somefield"

Compared to version 1.11, it comes from this addition in db/models/sql/compiler.py/get_combinator_sql():

if not compiler.query.values_select and self.query.values_select:
    compiler.query.set_values(self.query.values_select)

Here, self.query.values_select is ('id',).
For qs2, values_select is indeed empty, as coded in db/models/sql/query.py/set_values(),
because in this case the two values_list() arguments are managed in annotation_names[].

[Dimas Ari]

regression test result